Reconstructing impaired language using generative AI for people with aphasia

TL;DR

This paper explores using generative AI to help people with aphasia by reconstructing their impaired speech during conversations.

Contribution

A novel language-assistive solution using LLMs with in-context few-shot prompting to correct aphasic speech errors in real-time dialogue.

Findings

An AI solution using GPT-4o achieved 80% accuracy in reconstructing aphasic speech from a dataset of ~1980 utterances.

The system detects and corrects neologisms, paraphasic errors, and word-finding gaps in real-time conversations.

The Langchain architecture helped maintain conversation context and improve natural flow.

Abstract

In an era of Generative Artificial Intelligence (AI), it may be possible to capitalise on AI’s generative capabilities to assist people in compensating for their impaired language. Large Language Models (LLMs) have emerged as a recent breakthrough, revealing the potential to generate fluent, contextually relevant, and coherent texts. The current study leverages this inherent capability of LLMs in text generation and completion to compensate for impaired language in adults with acquired communication disabilities. To date, research studies on LLM for aphasia (a language-based communication disability after brain injury) have focused on specific and well-defined tasks and contexts (e.g., story retelling), and therefore may be less accurate and reliable in real-life conversation scenarios. This research proposes a language-assistive solution embedded in dialogue systems for individuals with aphasia to detect and correct errors in their aphasic speech during natural conversations. We have customised using in-context few-shot prompting (no weight updates) LLM to correct neologisms, paraphasic errors, and word-finding gaps that occur in aphasic speech. This could assist in identifying such errors in conversation and suggest completions in fragmented sentences. We utilised the Langchain architecture to retain previous utterances in memory, enabling the preservation of context and maintaining a natural conversation flow. We utilised a dataset comprising ~ 1980 utterances from 180 participants from AphasiaBank, and the AI-reconstructed utterances achieved an accuracy of 80% using the GPT-4o model. We further investigated the impact of different speech errors on reconstruction accuracy to determine which errors affect the capability of LLMs to correct errors in impaired speech.

The online version contains supplementary material available at 10.1038/s41598-025-24725-x.

Figures